Abstract
BACKGROUND: Due to altered drug clearance, renal impairment necessitates drug dose adjustments to prevent toxicity or therapeutic failure, yet inappropriate dosing persists. The utility of AI tools (e.g., ChatGPT®, DeepSeek®) in supporting renal dose adjustments remains understudied. OBJECTIVE: Evaluate renal dose adjustment practices in hospitalized patients and compare AI models (ChatGPT®, DeepSeek®) against UpToDate® guidelines. METHOD: A prospective observational study (January-April 2024) included hospitalized patients with creatinine clearance <60 mL/min in a general medicine ward. Medication regimens were assessed against UpToDate® guidelines. Five AI models (ChatGPT® 3.5, ChatGPT® 4.0, ChatGPT® 5.0, DeepSeek®, DeepThink®) were tested using 348 tailored prompts; sensitivity, specificity, and accuracy were calculated. RESULTS: Renal impairment prevalence was 30.9 %. Of 1461 drug orders, 23.8 % (348) required adjustment, with 63.5 % (221/348) inappropriately dosed in 76.4 % (113/148) of patients. Errors included 134 (38.5 %) unadjusted, 75 (21.6 %) inappropriately adjusted, and 12 (3.4 %) contraindicated regimens. Piperacillin/tazobactam (66.7 %), levofloxacin (83.33 %), and ranitidine (89.1 %) were most frequently inappropriately dosed. Severe renal impairment (CrCl ≤30 mL/min) increased improper dosing risk (AOR: 3.34; p = 0.004). DeepThink® showed the highest sensitivity (81.6 %) but low specificity (29 %) and 67.5 % accuracy. With advances in the ChatGPT® model, there was a modest improvement in prediction capacity, with the latest ChatGPT® 5.0 achieving balanced performance (64.3 % sensitivity, 54.8 % specificity, 61.7 % accuracy). CONCLUSION: Inappropriate renal dosing is prevalent among hospitalized patients with renal impairment. While AI models show promise as clinical decision support tools, their accuracy limitations require further optimization via evidence-based database training, prompt refinement, and integration of clinical expertise for reliable implementation.